AG-UI Output Format

infer headless can emit its stdout stream as AG-UI protocol events instead of the legacy CLI-specific JSON lines:

infer headless --format ag-ui "fix the failing test"

The default (--format json) is unchanged. With ag-ui, stdout carries exclusively newline-delimited AG-UI events, serialized with the official AG-UI Go SDK (github.com/ag-ui-protocol/ag-ui/sdks/community/go/pkg/core/events), so any AG-UI client can decode them with events.EventFromJSON without a custom adapter. AG-UI is transport agnostic, so a subprocess host reading stdout is a fully valid transport.

Event mapping

Agent lifecycle moment	AG-UI event(s)
Session start	RUN_STARTED with threadId = session id, runId = fresh per-invocation id
Resume via --session-id	MESSAGES_SNAPSHOT of the restored history, right after RUN_STARTED
User / assistant message	TEXT_MESSAGE_START / TEXT_MESSAGE_CONTENT / TEXT_MESSAGE_END, role on START
Assistant tool call	TOOL_CALL_START / TOOL_CALL_ARGS / TOOL_CALL_END (parented to the assistant message when it has text)
Tool result	TOOL_CALL_RESULT whose content is the raw JSON execution result (no "Result of tool call:" prefix)
Todo-list change	STATE_SNAPSHOT with {"todos": [...]}
Approval request (--require-approval)	CUSTOM event named approval_request carrying the legacy payload
Successful exit	RUN_FINISHED with a success outcome; result carries the session stats (tokens, cost)
Failure or panic	RUN_ERROR with the error message and the run id

Every run is bracketed by RUN_STARTED and exactly one terminal RUN_FINISHED or RUN_ERROR. The approval_request value is the legacy payload (tool_name, tool_args, tool_call_id); replies are still approval_response JSON lines on stdin, exactly as in json mode. Whole messages are emitted as single-delta triads (the synchronous headless loop produces complete messages); live token streaming is a planned follow-up.

Desktop sidecar wiring

A Tauri (or any subprocess-hosting) app consumes the stream by spawning the CLI as a sidecar and feeding each stdout line to its AG-UI client:

import { Command } from "@tauri-apps/plugin-shell";

const cmd = Command.sidecar("binaries/infer", [
  "agent", "--output-format", "ag-ui", task,
]);
cmd.stdout.on("data", (line) => {
  const event = JSON.parse(line); // an AG-UI BaseEvent, e.g. { type: "RUN_STARTED", ... }
  dispatchAgUiEvent(event);      // feed your AG-UI client / transformer
});
await cmd.spawn();

To resume a thread, pass the threadId from RUN_STARTED back as --session-id on the next invocation; the new run starts with a MESSAGES_SNAPSHOT so the client can render prior history.